CN111426383A - Hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis through machine learning - Google Patents

Hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis through machine learning Download PDF

Info

Publication number
CN111426383A
CN111426383A CN202010303994.XA CN202010303994A CN111426383A CN 111426383 A CN111426383 A CN 111426383A CN 202010303994 A CN202010303994 A CN 202010303994A CN 111426383 A CN111426383 A CN 111426383A
Authority
CN
China
Prior art keywords
polarization
full
image
sparse basis
stokes parameters
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010303994.XA
Other languages
Chinese (zh)
Other versions
CN111426383B (en
Inventor
许廷发
樊阿馨
王茜
张宇寒
潘晨光
郝建华
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Chongqing Innovation Center of Beijing University of Technology
Original Assignee
Beijing Institute of Technology BIT
Chongqing Innovation Center of Beijing University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Institute of Technology BIT, Chongqing Innovation Center of Beijing University of Technology filed Critical Beijing Institute of Technology BIT
Priority to CN202010303994.XA priority Critical patent/CN111426383B/en
Publication of CN111426383A publication Critical patent/CN111426383A/en
Application granted granted Critical
Publication of CN111426383B publication Critical patent/CN111426383B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01J—MEASUREMENT OF INTENSITY, VELOCITY, SPECTRAL CONTENT, POLARISATION, PHASE OR PULSE CHARACTERISTICS OF INFRARED, VISIBLE OR ULTRAVIOLET LIGHT; COLORIMETRY; RADIATION PYROMETRY
    • G01J3/00—Spectrometry; Spectrophotometry; Monochromators; Measuring colours
    • G01J3/28—Investigating the spectrum
    • G01J3/2823—Imaging spectrometer
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01J—MEASUREMENT OF INTENSITY, VELOCITY, SPECTRAL CONTENT, POLARISATION, PHASE OR PULSE CHARACTERISTICS OF INFRARED, VISIBLE OR ULTRAVIOLET LIGHT; COLORIMETRY; RADIATION PYROMETRY
    • G01J3/00—Spectrometry; Spectrophotometry; Monochromators; Measuring colours
    • G01J3/28—Investigating the spectrum
    • G01J3/447—Polarisation spectrometry
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]

Landscapes

  • Physics & Mathematics (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)

Abstract

The invention discloses a hyperspectral full-polarization image compression reconstruction method for optimizing sparse bases by machine learning, which adopts a combination of a quarter-wave plate and a device with linear polarization characteristics to image an image on a detector, and realizes different full-polarization modulation modes by switching the fast axis angle of the quarter-wave plate and/or the transmission axis angle of the device with linear polarization characteristics; processing the full-polarization local image of any wave band by adopting the full-polarization modulation mode to obtain compressed information; and optimizing a sparse basis by adopting a particle swarm algorithm, wherein the optimized sparse basis enables the full-polarization local image reconstructed by utilizing the compressed information to approach to the original image. When the method is applied, the full-polarization modulation mode is adopted to perform polarization modulation on the hyperspectral full-polarization image to obtain compressed information, and the optimized sparse basis is utilized to obtain a reconstructed hyperspectral full-polarization image. By adopting the method and the device, the reconstruction of the four Stokes parameters of the hyperspectral image can be realized, and the reconstruction precision of the four Stokes parameters is improved.

Description

Hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis through machine learning
Technical Field
The invention belongs to the technical field of hyperspectral polarization imaging, and particularly relates to a hyperspectral full-polarization image compression reconstruction method for optimizing sparse bases through machine learning, which is used for realizing the reconstruction of hyperspectral images with full-stokes parameters.
Background
The hyperspectral full-polarization imaging technology is an advanced technology for acquiring four-dimensional data of a target scene, wherein the data comprises two-dimensional spatial information, one-dimensional spectral information and polarization information represented by four Stokes parameters. Compared with the traditional hyperspectral imaging method, the additional polarization information can represent surface roughness, conductivity, molecular distribution, material components and the like. The hyperspectral full-polarization imaging technology has the advantage of identifying fine surface features, and is widely applied to the fields of biomedicine, environmental monitoring, remote sensing, exploration of earth resources, target identification and the like. The Fourier transform theory-based polarization imaging has the defects of serious noise pollution, channel crosstalk, bandwidth limitation and the like, and the development of the compressive sensing theory provides a new technology and method for polarization imaging. However, a typical Coded Aperture Snapshot Spectral Polarization Imaging (CASSPI) system is complex in structure, and cannot obtain circular polarization information represented by a fourth stokes parameter, so that the application of a polarization Imaging technology in the fields of circular polarization 3D glasses, mobile phone display screens and the like is limited.
To overcome the above-mentioned problem of lack of circular polarization information, a Channel Compressive Imaging Spectrometer (CCISP) based on Compressive sensing theory is proposed to reconstruct a polarization spectrum image containing four stokes parameters. The CCISP system consists of a linear polaroid rotating wheel, two high-order retarders and a linear polaroid, wherein the rotating wheel is provided with the linear polaroids with the transmission axes of 0 degree, 45 degrees, 90 degrees and 135 degrees in four directions respectively. The system improves the modulation freedom degree of the polarization state through combination, thereby realizing the modulation of four Stokes parameters. However, the measurement of circular polarization information greatly increases the complexity of the system and the loss of optical energy, and in the process of reconstructing four stokes parameters, the sampling time is long and the information is redundant due to the selection limitation of sparse bases.
Therefore, on the basis of ensuring that four stokes parameters are obtained simultaneously, the full polarization imaging technology based on the compressive sensing theory needs to be further improved and developed. The complexity of the system is reduced, the cost can be reduced, and the operation flow of compression measurement can be simplified. In the reconstruction process, by increasing the sparse basis selection flexibility of the four Stokes parameters, the sampling rate can be reduced, the sparse basis selection process can be omitted, and the reconstruction precision can be improved.
Disclosure of Invention
In view of the above, the invention provides a hyperspectral full-polarization image compression reconstruction method for optimizing sparse bases through machine learning, which can realize the reconstruction of four stokes parameters of a hyperspectral image and improve the reconstruction accuracy of the four stokes parameters.
In order to solve the technical problem, the invention is realized as follows:
a hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis by machine learning comprises the following steps:
step 1, imaging an image on a detector by adopting a combination of a quarter-wave plate and a device with linear polarization characteristics, and realizing different full-polarization modulation modes by switching a fast axis angle of the quarter-wave plate and/or a transmission axis angle of the device with linear polarization characteristics, wherein the switching times are less than or equal to 3; by adopting the full-polarization modulation mode, four Stokes parameters F of each spatial pixel of each waveband in the hyperspectral full-polarization image F to be reconstructed are subjected to polarization modulation, and compressed information g is obtained*The compressed information G can be obtained by combining all spatial pixels of all bands*；
Step 2, constructing the Mueller matrix by utilizing the quarter-wave plate and the device with the linear polarization characteristicTaking the first row of elements of the polarization modulation matrix H, and substituting the fast axis angle and the transmission axis angle used for detection in the step 1 to obtain the polarization modulation matrix H*；
Step 3, knowing the full-polarization local image F of any lambda wave band in the high-spectrum full-polarization image F to be reconstructedλExtracting a full-polarization partial image FλFour stokes parameters f of each spatial pixelλ(ii) a Using polarization modulation matrix H*Four Stokes parameters f for each spatial pixelλPolarization modulation is carried out to obtain compressed information gλ *Full polarization partial image FλAll spatial pixels are combined to obtain compressed information Gλ *；
Step 4, sparse basis psi of four Stokes parameters is subjected to particle swarm optimization*Performing iterative optimization; wherein elements in the particles correspond to elements in the sparse basis matrix; after each iteration, aiming at each obtained sparse basis, based on a compressed sensing theory, utilizing the sparse basis and the polarization modulation matrix H*And the compression information g obtained in step 3λ *Reconstructing a full-polarization partial image FλFour Stokes parameters f per spatial pixelλ *All spatial pixels are combined to obtain a reconstructed full-polarization partial image Fλ *(ii) a Using known fully-polarized partial images FλAnd reconstructed full polarization partial image Fλ *Calculating peak signal-to-noise ratios (PSNR) of four Stokes parameters of the reconstructed local image, and recording the minimum value of the four PSNR as an individual optimal value; updating the individual optimal value and the individual optimal particle position on the basis of keeping a larger value, and keeping the maximum value of the individual optimal value as a new global optimal value; when the preset optimization stopping condition is reached, the final global optimal value g corresponds to the global optimal particle position xgTaking the sparse basis corresponding to the global optimal particle position as polarization modulation H*Corresponding full stokes parameter sparse basis Ψ*；
Step 5, based on the compressed sensing theory, adopting the optimized sparse basis psi*Polarization modulation matrix H*And the compression information g obtained in step 1*Reconstructing four Stokes parameters f of each spatial pixel of each band*Combining all spatial pixels of all bands to obtain high-spectrum full-polarization image F*。
Wherein the device having linear polarization characteristics is a linear polarizer or a liquid crystal tunable filter (L CTF).
Has the advantages that:
(1) according to the invention, the particle swarm algorithm calculation is carried out by adopting the full-polarization local image of any wave band, so that an optimal sparse basis can be obtained, and the sparse basis can be used for reconstructing a high-spectrum full-polarization image. The full-polarization sparse basis is optimized by utilizing the particle swarm algorithm, the sparse basis selection problem in the compressive sensing theory is avoided, and the reconstruction precision of the four Stokes parameters is improved. The optimization pertinence is strong for different polarization modulation modes, the adaptability of the sparse basis is good, and a more appropriate polarization modulation mode can be selected according to target requirements, namely the fast axis angle of the quarter-wave plate and the transmission axis angle of a device with linear polarization characteristics.
(2) The full polarization compression is performed by combining the quarter-wave plate and the device with the linear polarization characteristic, and the number of devices for polarization modulation is reduced.
(3) The sparse basis is optimized by adopting the particle swarm algorithm, the sparse basis of the four Stokes parameters at least consists of 16 elements, namely the particle dimension in the particle swarm algorithm is at least 16, and the optimization speed of the algorithm is high because only the four Stokes parameters of the local image need to be reconstructed. The scheme can provide a new idea for sparse basis selection when a compressive sensing theory is applied in more fields.
Drawings
FIG. 1 is a flow chart of a machine learning optimization sparse basis hyperspectral full-polarization image compression reconstruction.
FIG. 2 is a diagram of an example of a machine learning optimized sparse basis hyperspectral full-polarization image compression reconstruction.
Fig. 3 is a schematic view of a combination of a quarter-wave plate and a linear polarizer.
FIG. 4 is a graph of PSNR minimum values of four Stokes parameters of a particle swarm optimization sparse basis and a reconstructed 600nm waveband local image along with the change of iteration times.
Fig. 5 is a PSNR diagram of a reconstructed hyperspectral fully polarized image.
Detailed Description
The invention is described in detail below by way of example with reference to the accompanying drawings.
The invention provides a hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis by machine learning, and the flow is shown in figure 1. The method is realized by two steps: obtaining sparse bases of four Stokes parameters based on a compressed sensing theory and a machine learning algorithm according to a full-polarization local image of any known wave band; and reconstructing four Stokes parameters of the hyperspectral full-polarization image based on a compressed sensing theory and an optimized sparse basis according to the hyperspectral full-polarization image to be detected.
And according to the known full-polarization local image of any wave band, performing compression reconstruction of a full-Stokes parameter in a certain polarization modulation mode, and optimizing sparse basis. And compressing the full-polarization global image of the waveband in the same polarization modulation mode. The reconstruction effect is determined by the finite equidistant Property (RIP) of the sensing matrix (composed of the polarization modulation matrix and the sparse basis matrix), and has little relation with the full stokes parameters for compression. Because the same polarization modulation mode is adopted, the optimized sparse basis can be used for reconstructing the full Stokes parameters of the full-polarization global image of the waveband. In the same way, under the same polarization modulation mode, the optimized sparse basis can be used for reconstructing the full stokes parameters of full polarization images of other wave bands.
The following describes in detail the scheme and the result of the particle swarm optimization full stokes parameter sparse basis and the hyper-spectral image full stokes parameter reconstruction based on a group of hyper-spectral full polarization image data sets and a randomly selected full polarization modulation mode, as shown in fig. 2-5.
Step 1, optimizing full Stokes parameter sparse basis by particle swarm optimization
Step 11, defining the hyperspectral full-polarization image to be measured as
Figure BDA0002455072880000051
Wherein, 4 of the upper corner mark is four Stokes parameters of the polarization dimension of the hyperspectral full-polarization image, NλNumber of bands, N, for spectral dimensions of a hyperspectral polarimetric imagex×NyThe number of pixels in the spatial dimension of the hyperspectral full-polarization image. Meanwhile, the invention is premised on that the full-polarization local image of any lambda wave band is known
Figure BDA0002455072880000052
(Mx<Nx，My<Ny)，Mx×MyThe number of pixels in the spatial dimension of the fully polarized local image. Then the full polarization partial image F can be extractedλFour stokes parameters f of each spatial pixelλ。
In this embodiment, as shown in FIG. 2, a set of hyperspectral full-polarization image datasets is known
Figure BDA0002455072880000053
Including four Stokes parameters, the central wavelengths of 18 spectral bands are from 520nm to 690nm at intervals of 10nm, the space dimension is 800 × 800 pixels, and a full-polarization partial image of 600nm band is taken
Figure BDA0002455072880000054
Including four Stokes parameters, with a spatial dimension of 200 × 200 pixels, and the four Stokes parameters for each spatial pixel are recorded as
Figure BDA0002455072880000055
High spectral full polarization image dataset in the present embodiment
Figure BDA0002455072880000056
And verifying the reconstruction result. In practical application, F is a hyperspectral full-polarization image to be reconstructed and is an unknown image to be reconstructed. The invention only needs to obtain the full-polarization local image F of a certain lambda wave bandλThen is ready toSo as to obtain an optimized sparse basis through a particle swarm algorithm. The sparse basis can be used for reconstructing a full-polarization image of the lambda wave band and reconstructing full-polarization images of other wave bands, so that a high-spectrum full-polarization image is obtained. The reconstruction methods of the full-polarization images of different wave bands are the same and are based on a compressed sensing theory.
And step 12, constructing a combination of the quarter-wave plate and the linear polarizer, as shown in fig. 3, wherein along the light propagation direction, the light wave is subjected to polarization modulation by the quarter-wave plate and the linear polarizer in sequence, and then is imaged on the detector. The quarter-wave plate is marked with a fast axis direction, and the fast axis angle is an included angle between the fast axis and the horizontal direction and can be switched between 0 degree and 180 degrees; the linear polarizer is marked with a light transmission axis direction, and the light transmission axis angle is an included angle between the light transmission axis and the horizontal direction and can be switched from 0 degree to 180 degrees. In polarization modulation, a fast axis angle of the quarter-wave plate and a transmission axis angle of the linear polarizer are combined to obtain polarization modulation information. To obtain a set of polarization modulation information, the transmission axis angle of the linear polarizer can be fixed, and the set of fast axis angles of the quarter-wave plate can be switched sequentially.
Step 13, constructing a polarization modulation matrix H by using the Mueller matrix of the quarter-wave plate and the linear polarizer, and taking the first row of elements to further obtain the polarization modulation matrix
Figure BDA0002455072880000061
(n is more than or equal to 1 and less than or equal to 3). Here, n is equal to or less than 3 to realize compression, and n indicates that the angle combination of the quarter-wave plate and the linear polarizer is switched several times. In practice, a quarter-wave plate or linear polarizer may be fixed and a set of full polarization compression results obtained by changing the angle of the other device.
In this embodiment, the Mueller matrix of the quarter-wave plate is
Figure BDA0002455072880000062
The Mueller matrix of the linearly polarizing plate is
Figure BDA0002455072880000063
Thus, the polarization modulation matrix of the quarter-wave plate and linear polarizer combination is H ═ M2×M1，
Figure BDA0002455072880000064
Figure BDA0002455072880000065
Figure BDA0002455072880000071
Figure BDA0002455072880000072
In the formula, theta is an included angle between a fast axis of the quarter-wave plate and the horizontal direction, α is an included angle between a transmission axis of the linear polarizer and the horizontal direction, and the detector can only receive total light intensity information, namely a first Stokes parameter, corresponding to a first row element of a polarization modulation matrix H.
In the embodiment, the fast axis angle of the quarter-wave plate is selected to be 5 degrees and 144 degrees in sequence, the transmission axis of the linear polarizer is always horizontal (α is 0), and (theta is 5 degrees, α is 0 degrees) and (theta is 144 degrees, α is 0 degrees) are respectively substituted into the first row of elements of H to obtain two rows of elements, and the two rows of elements are combined to obtain a polarization modulation matrix
Figure BDA0002455072880000073
The degrees of theta and α can be chosen at will, so long as H is constructed*The selected angle of the compressed image is the same as the angle of the compressed image acquired in step 21.
Figure BDA0002455072880000074
Step 14, using polarization modulation matrix H*For full polarization localImage FλFour stokes parameters f of each spatial pixelλPolarization modulation is carried out to obtain compressed information gλ *。
In this step, a polarization modulation matrix H is used*Carrying out polarization modulation on four Stokes parameters of each spatial pixel in a 600nm wave band full-polarization local image to obtain compression information gλ *＝H*×fλ，
Figure BDA0002455072880000075
Compression information g for all spatial pixelsλ *Compressed information of four Stokes parameters of 600nm wave band full polarization partial image obtained by combination
Figure BDA0002455072880000076
Including two polarization modulating fast axis angles of 5 degrees and 144 degrees, the spatial dimension is 200 × 200 pixels.
The following steps 15-17 are sparse radix Ψ of four Stokes parameters by adopting a particle swarm optimization*And performing iterative optimization. Wherein the elements in the particles correspond to elements in the sparse basis matrix. Particle dimension D ═ B1×B2The sparse radical being B1×B2(B1＝4，B2≧ 4). The transformation relation between the particles and the sparse basis is that D-dimensional data of the particles are utilized to construct B1×B2Of the matrix of (a). The positional correspondence of the element in the particle and in the sparse basis is not necessarily limited. For example, a way of correspondence between particle elements and sparse base elements is as follows: to increase the optimization rate of the algorithm, take B1＝B2With 4. D. 16, the first column with 1-4 elements as the sparse base, the 2 nd column with 5-8 elements as the sparse base, and so on, a sparse base of 4 × 4 is obtained.
Step 15, setting parameters of the particle swarm algorithm as follows: the number of particles N is 20, the particle dimension D is 16, the maximum number of iterations T is 100, and a learning factor c1＝c21.5, maximum value of inertial weight Wmax0.8, minimum value of inertial weight Wmin0.4, position max X max1, minimum value of position XminMaximum value of speed V ═ 1max0.01, minimum speed Vmin-0.01. The initial value of the position of the population is
Figure BDA0002455072880000081
Each value is a position interval [ -1,1 [ ]]The random number of (1). The initial value of the population velocity is
Figure BDA0002455072880000082
Each value is a speed interval [ -0.01,0.01 [ -0.01 [ ]]The random number of (1). The initialized values can be adjusted according to the needs, but are not necessary.
The following is the process of full stokes parametric reconstruction with one particle in the population as the sparse basis.
Initial value of position of a particle in a population
Figure BDA0002455072880000083
Deformation (x)0I.e. the initial value of the particle, the deformation operation is the deformation from the vector to the matrix realized according to the corresponding relation) to obtain the sparse basis of the full stokes parameter
Figure BDA0002455072880000084
Four stokes parameters for each spatial pixel
Figure BDA0002455072880000085
Performing sparse representation fλ＝Ψ*×θ，
Figure BDA0002455072880000086
A sparse vector of full stokes parameters. Thus, the polarization modulation of the four stokes parameters for each spatial pixel can be expressed as gλ *＝H*×fλ＝H*×Ψ*×θ＝A*×θ，A*＝H*×Ψ*Is a perceptual matrix.
In the known polarization modulation matrix H*Four stokes parameters per spatial pixelCompressed information g of volumeλ *And sparse base Ψ*In the case of (2), the sparse vector θ can be solved*Four Stokes parameters f of each spatial pixel are reconstructedλ *＝Ψ*×θ*The four Stokes parameters of all the spatial pixels are combined to obtain a full-polarization local image with a 600nm wave band
Figure BDA0002455072880000087
Sparse vector θ*This can be obtained by solving the following optimization problem:
Figure BDA0002455072880000091
in the formula, tau is l1Norm regularization weights. The value τ in the reconstruction process shown in fig. 2 is 0.005, and the optimization problem can be solved by using a Two-Step Iterative Shrinkage/threshold (TwIST-Step Iterative Shrinkage/threshold, TwIST) algorithm.
Each particle in the population can be reconstructed as a sparse basis using a known full-polarization partial image FλAnd reconstructed full polarization partial image Fλ *Calculating peak signal-to-Noise Ratio (PSNR) of four Stokes parameters of the reconstructed local image, recording the minimum value of the four PSNR values as an individual optimal value p, and simultaneously recording the corresponding particle position
Figure BDA0002455072880000092
(i.e., the value of the particle, i.e., the rarefaction basis), each particle has its own p. The individual optimal values of 20 particles in the population are recorded as vectors
Figure BDA0002455072880000093
Individual optimal particle locations are noted
Figure BDA0002455072880000094
Recording the maximum value in the individual optimal values p of 20 particles in the population as a global optimal value g, and simultaneously recording the corresponding valuesGlobal optimal particle position
Figure BDA0002455072880000095
Step 16, in the ith iteration process of the particle swarm optimization, the position X of the population is updated by using the dynamic inertia weight and the speedi. And obtaining 20 sparse bases at each iteration, and calculating the minimum value of the four Stokes parameters PSNR of the reconstructed local image for each sparse base. The minimum value is taken for comparison in order to find a sparse basis in which the minimum value is also large, that is, the PSNR values of the four stokes parameters are integrally high. Comparison of XiThe PSNR minimum value calculated by each particle in the particle and the individual optimal value of the particle are updated by keeping the larger value as the principle
Figure BDA0002455072880000096
And individual optimal particle position
Figure BDA0002455072880000097
(for the t-th particle, compare xitPSNR minimum and individual optimum ptUpdating the larger value as a new individual optimal value to p, and updating the position of the particle corresponding to the new individual optimal value to XpMiddle), and updates the global optimal value g and the global optimal particle position
Figure BDA0002455072880000098
Step 17, when the preset optimization stop condition is reached, the final global optimal value g corresponds to the global optimal particle position xgTaking a sparse basis obtained by deforming the global optimal particle position as polarization modulation H*Corresponding full stokes parameter sparse basis Ψ*。
The particle swarm optimization shown in fig. 2 does not preset a threshold of the global optimum value in the sparse basis iterative optimization process, so that the optimization is stopped when a preset maximum iteration number is reached, and a specific result in the iterative process is shown in fig. 4. Finally obtaining the global optimal value g which is 24.16dB and the global optimal particle position
Figure BDA0002455072880000101
Deforming to obtain sparse basis
Figure BDA0002455072880000102
I.e. polarization modulation H*The corresponding full stokes parameter sparse basis.
Figure BDA0002455072880000103
Obtaining full-polarization local image of 600nm wave band based on sparse basis weight structure
Figure BDA0002455072880000104
Reconstructed full-polarization local image Stokes parameter S0、S1、S2、S3The PSNR values of the two-dimensional image are 33.86dB, 24.24dB, 24.16dB and 24.24dB in sequence.
Step 2, full Stokes parameter reconstruction of hyperspectral image
Step 21, imaging the image on a detector by adopting the combination of the quarter-wave plate and the device with the linear polarization characteristic, and realizing different full polarization compression by switching the fast axis angle of the quarter-wave plate and/or the transmission axis angle of the device with the linear polarization characteristic to obtain compression information g of four Stokes parameters F of each spatial pixel of each waveband in the hyperspectral full-polarization image F to be reconstructed*. Since images of all bands need to be acquired to form a hyperspectral image, this step needs to acquire compressed information g of four stokes parameters F of each spatial pixel of all bands in the hyperspectral full-polarization image F to be reconstructed*。
Since in this example, construction H*And optimizing sparse radix Ψ*When the fast axis angle of the quarter-wave plate is selected to be 5 degrees and 144 degrees in sequence, the transmission axis of the linear polarizer is always horizontal (α is equal to 0), so that the same angle combination is adopted when full-polarization compression imaging is actually carried out on the hyperspectral full-polarization image F
Figure BDA0002455072880000105
Including two polarization modulated fast axis angles of 5 degrees and 144 degrees, with 18 spectral bands centered at wavelengths spaced 10nm from 520nm to 690nm, and with a spatial dimension of 800 × 800 pixels, since here the information for all bands is obtained when obtaining the full polarization compressed information, a hyperspectral full polarization image for all bands can be reconstructed in step 22.
Step 22, utilizing the full Stokes parameter sparse base obtained in step 17
Figure BDA0002455072880000111
Based on the compressed sensing theory in step 15, the polarization modulation matrix H obtained in step 13 is utilized*And the compression information g obtained in step 21*Four Stokes parameters f of each spatial pixel of each waveband in reconstructed high spectrum full polarization image*And further obtaining a reconstructed hyperspectral full-polarization image F*。
In the reconstruction process, the four Stokes parameters of each spatial pixel of each waveband in the hyperspectral full-polarization image are subjected to polarization modulation and can be expressed as g*＝H*×f＝H*×Ψ*×θ＝A*×θ，A*＝H*×Ψ*Is a perceptual matrix. H*As is known, Ψ*For the optimized result of the particle swarm optimization, g*From step 21, θ can be calculated by substituting the above equation*. Four Stokes parameters f of each spatial pixel of each band are then reconstructed*＝Ψ*×θ*. Then, the four Stokes parameters of all spatial pixels of all bands are combined to obtain a reconstructed hyperspectral fully-polarized image
Figure BDA0002455072880000112
Through the reconstruction in the step 22, the Stokes parameter S of the reconstructed hyperspectral full-polarization image in the 600nm wave band0、S1、S2、S3The PSNR values are 40.23dB, 27.52dB, 25.62dB and 26.92dB in sequence, and the reconstructed hyperspectral image is fullStokes parameter S of polarization image0、S1、S2、S3The average PSNR values in 18 spectral bands are 39.22dB, 23.99dB, 22.78dB and 23.29dB in sequence. As can be seen from the reconstruction results shown in fig. 5, the present embodiment successfully reconstructs the four stokes parameters of the hyperspectral full-polarization image.
In conclusion, the invention provides a hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis by machine learning. According to the known full-polarization local image of any wave band, based on a compressed sensing theory and a particle swarm algorithm, sparse bases of four Stokes parameters are successfully obtained. According to the hyperspectral full-polarization image to be measured, four Stokes parameters of the hyperspectral full-polarization image are successfully reconstructed based on a compressed sensing theory and an optimized sparse basis.
The linear polarizer in the above embodiment may be replaced by a liquid crystal Tunable Filter (L i required crystal Tunable Filter, L CTF) having linear polarization characteristics, and since L CTF has high spectral resolution characteristics, full polarization compression and high spectral imaging are realized at the same time.
In summary, the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (2)

1. A method for compressing and reconstructing a sparse basis hyperspectral full-polarization image by machine learning optimization is characterized by comprising the following steps:
step 1, imaging an image on a detector by adopting a combination of a quarter-wave plate and a device with linear polarization characteristics, and realizing different full-polarization modulation modes by switching a fast axis angle of the quarter-wave plate and/or a transmission axis angle of the device with linear polarization characteristics, wherein the switching times are less than or equal to 3; by adopting the full-polarization modulation mode, four Stokes parameters F of each spatial pixel of each waveband in the hyperspectral full-polarization image F to be reconstructed are subjected to polarization modulation, and compressed information g is obtained*All spatial pixels of all bands are combined to obtainCompressed information G*；
Step 2, constructing a polarization modulation matrix H by utilizing the quarter-wave plate and the Mueller matrix of the device with the linear polarization characteristic, taking the first row of elements, and substituting the fast axis angle and the transmission axis angle for detection in the step 1 to obtain the polarization modulation matrix H*；
Step 3, knowing the full-polarization local image F of any lambda wave band in the high-spectrum full-polarization image F to be reconstructedλExtracting a full-polarization partial image FλFour stokes parameters f of each spatial pixelλ(ii) a Using polarization modulation matrix H*Four Stokes parameters f for each spatial pixelλPolarization modulation is carried out to obtain compressed information gλ *Full polarization partial image FλAll spatial pixels are combined to obtain compressed information Gλ *；
Step 4, sparse basis psi of four Stokes parameters is subjected to particle swarm optimization*Performing iterative optimization; wherein elements in the particles correspond to elements in the sparse basis matrix; after each iteration, aiming at each obtained sparse basis, based on a compressed sensing theory, utilizing the sparse basis and the polarization modulation matrix H*And the compression information g obtained in step 3λ *Reconstructing a full-polarization partial image FλFour Stokes parameters f per spatial pixelλ *All spatial pixels are combined to obtain a reconstructed full-polarization partial image Fλ *(ii) a Using known fully-polarized partial images FλAnd reconstructed full polarization partial image Fλ *Calculating peak signal-to-noise ratios (PSNR) of four Stokes parameters of the reconstructed local image, and recording the minimum value of the four PSNR as an individual optimal value; updating the individual optimal value and the individual optimal particle position on the basis of keeping a larger value, and keeping the maximum value of the individual optimal value as a new global optimal value; when the preset optimization stopping condition is reached, the final global optimal value g corresponds to the global optimal particle position xgTaking the sparse basis corresponding to the global optimal particle position as polarization modulation H*Corresponding full stokes parameter sparse basis Ψ*；
Step 5, based on the compressed sensing theory, adopting the optimized sparse basis psi*Polarization modulation matrix H*And the compression information g obtained in step 1*Reconstructing four Stokes parameters f of each spatial pixel of each band*Combining all spatial pixels of all bands to obtain high-spectrum full-polarization image F*。
2. The method of claim 1, wherein the device having linear polarization characteristics is a linear polarizer or a liquid crystal tunable filter (L CTF).
CN202010303994.XA 2020-04-17 2020-04-17 Hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis through machine learning Active CN111426383B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010303994.XA CN111426383B (en) 2020-04-17 2020-04-17 Hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis through machine learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010303994.XA CN111426383B (en) 2020-04-17 2020-04-17 Hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis through machine learning

Publications (2)

Publication Number Publication Date
CN111426383A true CN111426383A (en) 2020-07-17
CN111426383B CN111426383B (en) 2022-05-10

Family

ID=71553978

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010303994.XA Active CN111426383B (en) 2020-04-17 2020-04-17 Hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis through machine learning

Country Status (1)

Country Link
CN (1) CN111426383B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111998945A (en) * 2020-08-21 2020-11-27 北京理工大学重庆创新中心 Full-polarization hyperspectral image compression reconstruction method based on Stokes parameter blocking
CN112781728A (en) * 2020-12-30 2021-05-11 北京理工大学重庆创新中心 Full-polarization hyperspectral imaging method for accurately solving combined compressed sensing
CN115632761A (en) * 2022-08-29 2023-01-20 哈尔滨工业大学（深圳） Multi-user distributed privacy protection regression method and device based on secret sharing
CN117571128A (en) * 2024-01-16 2024-02-20 长春理工大学 High-resolution polarized spectrum image imaging method and system

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120069170A1 (en) * 2010-07-11 2012-03-22 Spynsite, LLC System for image rendering or spectral recognition
US20140353475A1 (en) * 2007-12-06 2014-12-04 U.S. Army Research Laboratory System and processor implemented method for improved image quality and enhancement based on quantum properties
US20160258645A1 (en) * 2015-03-06 2016-09-08 Mitsubishi Electric Research Laboratories, Inc. System and Method for Controlling Operations of Air-Conditioning System

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140353475A1 (en) * 2007-12-06 2014-12-04 U.S. Army Research Laboratory System and processor implemented method for improved image quality and enhancement based on quantum properties
US20120069170A1 (en) * 2010-07-11 2012-03-22 Spynsite, LLC System for image rendering or spectral recognition
US20160258645A1 (en) * 2015-03-06 2016-09-08 Mitsubishi Electric Research Laboratories, Inc. System and Method for Controlling Operations of Air-Conditioning System

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111998945A (en) * 2020-08-21 2020-11-27 北京理工大学重庆创新中心 Full-polarization hyperspectral image compression reconstruction method based on Stokes parameter blocking
CN111998945B (en) * 2020-08-21 2022-11-01 北京理工大学重庆创新中心 Full-polarization hyperspectral image compression reconstruction method based on Stokes parameter blocking
CN112781728A (en) * 2020-12-30 2021-05-11 北京理工大学重庆创新中心 Full-polarization hyperspectral imaging method for accurately solving combined compressed sensing
CN112781728B (en) * 2020-12-30 2024-03-19 北京理工大学重庆创新中心 Full-polarization hyperspectral imaging method capable of accurately solving and combining compressed sensing
CN115632761A (en) * 2022-08-29 2023-01-20 哈尔滨工业大学（深圳） Multi-user distributed privacy protection regression method and device based on secret sharing
CN115632761B (en) * 2022-08-29 2023-06-23 哈尔滨工业大学（深圳） Multi-user distributed privacy protection regression method and device based on secret sharing
CN117571128A (en) * 2024-01-16 2024-02-20 长春理工大学 High-resolution polarized spectrum image imaging method and system
CN117571128B (en) * 2024-01-16 2024-03-29 长春理工大学 High-resolution polarized spectrum image imaging method and system

Also Published As

Publication number Publication date
CN111426383B (en) 2022-05-10

Similar Documents

Publication Publication Date Title
CN111426383B (en) Hyperspectral full-polarization image compression reconstruction method for optimizing sparse basis through machine learning
CN108955887B (en) LCTF-based full-polarization hyperspectral compressive sensing imaging method
Marwah et al. Compressive light field photography using overcomplete dictionaries and optimized projections
US7427932B2 (en) Compressive sampling and signal inference
US20220108466A1 (en) System and method for reconstruction of compressed signal data using artificial neural networking
CN108955882B (en) Three-dimensional data reconstruction method based on liquid crystal hyperspectral calculation imaging system
Han et al. Tensor FISTA-Net for real-time snapshot compressive imaging
CN116245726A (en) Compressed sensing polarization super-resolution imaging method based on deep learning framework
CN109682476B (en) Method for performing compressed hyperspectral imaging by using self-adaptive coding aperture
CN113008370A (en) Three-dimensional self-adaptive compression reconstruction method based on liquid crystal hyperspectral calculation imaging system
CN114739511A (en) Snapshot type hyperspectral imaging chip structure based on spectrum modulation array
CN111998945B (en) Full-polarization hyperspectral image compression reconstruction method based on Stokes parameter blocking
CN112113661B (en) Deep learning type snapshot spectrum imaging device and detection method thereof
CN109781260B (en) Ultra-compact snapshot type polarization spectrum imaging detection device and detection method
CN107580164A (en) The compressed sensing super-resolution imaging system and its imaging method of a kind of single measurement
Yin et al. A deep network based on wavelet transform for image compressed sensing
CN109798978B (en) Spectrum imaging method adopting special-shaped blue noise coding aperture
CN116609942A (en) Sub-aperture compressed sensing polarization super-resolution imaging system and method
CN115824412A (en) Full-polarization spectrum imaging device and detection method
CN116797676A (en) Method suitable for coded aperture compression spectrum polarization imaging reconstruction
Han et al. Deep image prior plus sparsity prior: toward single-shot full-Stokes spectropolarimetric imaging with a multiple-order retarder
CN112781728B (en) Full-polarization hyperspectral imaging method capable of accurately solving and combining compressed sensing
Shen et al. Development of energy correction based multishot snapshot spectral imaging system
Zhang et al. Bi-channel compressive hyperspectral imager based on polarization conversion metasurface
Zhang et al. Snapshot compressive hyperspectral imaging via polarization conversion metasurface

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant